Modern generative pipelines churn out thousands of de novo protein candidates with ease, yet wet-lab validation remains an expensive, painfully slow chokepoint. In research accepted at the ICML 2026 Workshop on Generative and Agentic AI for Biology, Gyubok Lee, Kiwoong Yoo, Jimin Seo, Kyunghoon Hur, and Edward Choi introduce a post-generation decision layer that deploys LLMs to build multi-metric ranking and shortlisting policies before running costly bench experiments.
Rather than relying on arbitrary manual filtering, the framework prompts language models to synthesize ranking logic over heterogeneous confidence scores and interface proxies generated by tools like Boltz-2 and Protenix. The authors report that averaging five sampled global iterative gpt-4o policies delivered 0.589 Recall@10 on a 10-target held-out benchmark, surpassing the single-feature Protenix binder ipTM baseline of 0.571 Recall@10.
On a targeted subset covering Nipah, RBX1, and TREM2, target-conditioned iterative policies reached 0.519 Recall@10 and 0.583 NDCG@10. In our view, treating candidate evaluation as an interpretable meta-ranking task directly tackles the core economics of computational biology—weeding out non-viable binders early to slash physical synthesis waste and accelerate real-world discovery.